ObjectiveTo compare the diagnostic performance of contrast-enhanced transcranial Doppler (c-TCD) and contrast-enhanced transthoracic echocardiography (c-TTE) under resting and Valsalva conditions for detecting patent foramen ovale (PFO), and to evaluate the clinical utility of an integrated diagnostic model combining both modalities.Materials and methodsIn this retrospective study, a total of 146 patients with suspected PFO underwent both c-TCD and c-TTE at rest and during the Valsalva maneuver. Among them, 40 patients also received contrast-enhanced transesophageal echocardiography (c-TEE), which served as the reference standard. Detection rates, shunt grading, and inter-modality agreement were analyzed. A logistic regression model integrating c-TCD and c-TTE findings was developed and assessed using receiver operating characteristic (ROC) curves and decision curve analysis (DCA).ResultsThe Valsalva maneuver significantly increased the detection rates of right-to-left shunt (c-TCD: from 41.4% to 76.7%; c-TTE: from 45.2% to 80.1%) and the proportion of Grade 3 shunts (P < 0.0001). At rest, c-TCD detected more high-grade shunts than c-TTE (P = 0.030), supporting its use as a sensitive initial screening tool. Inter-modality agreement improved markedly during Valsalva (weighted Kappa = 0.782). The combined model achieved 100% sensitivity and negative predictive value in the TEE subgroup, with an area under the ROC curve of 0.97 and no false negatives. DCA confirmed the superior net clinical benefit of the integrated approach across a broad range of decision thresholds.ConclusionValsalva maneuver significantly enhances the diagnostic yield of both c-TCD and c-TTE. While c-TCD may serve as an effective first-line screening tool, its combination with c-TTE ensures improved diagnostic accuracy and clinical decision-making value. The integrated model demonstrates strong potential for clinical implementation as a noninvasive, efficient strategy to reduce unnecessary c-TEE procedures.

To develop and validate a combined ultrasound-based radiomics-clinical model for differentiating benign and malignant breast lesions. A total of 3142 patients from eight hospitals between February 2012 and September 2024 were included in this multicenter retrospective development and validation study, with an additional single-center prospective test cohort. Lesions were manually segmented, and radiomics features were automatically extracted to construct five machine learning models. The best-performing radiomics model was combined with clinical features to build a combined model. Model performance and its impact on Breast Imaging Reporting and Data System (BI-RADS)-based biopsy decisions were evaluated. Logistic regression (LR) showed the best radiomics performance, with area under the curves (AUCs) of 0.83, 0.82, 0.81, and 0.82 across the training, internal test, external test, and prospective test sets. The clinical model achieved AUCs of 0.87, 0.85, 0.87, and 0.86, whereas the combined model achieved AUCs of 0.92, 0.90, 0.92, and 0.93, significantly outperforming both single-modality models (all p < 0.01). Decision curve analysis (DCA) showed that the combined model had a higher net benefit than the other models across a broad range of threshold probabilities (0.05–0.95) in this study. Performance remained stable across lesion size and age subgroups. In the reclassification analysis, the model suggested the potential to influence biopsy recommendations without a significant reduction in sensitivity and to increase the malignancy yield in BI-RADS 4a. Shapley additive explanations (SHAP) analysis provided clinically interpretable feature contributions. The interpretable ultrasound-based radiomics model enables reliable, noninvasive breast lesion diagnosis and may reduce unnecessary biopsies. This work developed an interpretable radiomics-clinical combined model in a multicenter retrospective development and validation study, with additional testing in a single-center prospective cohort, and may support breast lesion risk stratification and biopsy decision-making after further prospective clinical utility evaluation.

Ovarian cancer (OC) is a leading cause of gynecologic cancer mortality, with survival prediction limited by existing prognostic models that fail to capture tumor heterogeneity. Conventional methods lack precision for individualized risk assessment. Deep learning (DL) addresses these issues by integrating diverse data to improve survival prediction and risk stratification. This study introduces OvcaSurvivor, a novel multimodal DL framework for R0-resected OC patients, integrating whole-slide images (WSI), ultrasound (US), and clinical data from 543 patients. It uses advanced neural networks (CHIEF for WSI, ResNet50 for US) and an attention-guided fusion module. OvcaSurvivor showed superior performance, with C-indices of 0.81 (internal), 0.76 (external 1), and 0.70 (external 2). Time-dependent AUCs for 1-, 3-, and 5-year survival were highly accurate. WSI features drove prediction, and the model stratified patients into high/low-risk groups, highlighting clinical utility. This multimodal fusion advances OC precision oncology, enabling robust postoperative management.

BackgroundTo develop and validate a multimodal habitat radiomics model integrating automated breast volume scanning (ABVS) and conventional two-dimensional ultrasound (2D-US) for risk stratification of biopsy-selected BI-RADS 4A breast lesions.MethodsThis retrospective single-center study included 160 consecutive patients with BI-RADS 4A breast lesions confirmed by histopathology between January 2024 and May 2025. Tumoral and peritumoral regions were manually segmented on ABVS and 2D-US images. Habitat subregions were generated using a local spatial autocorrelation-based heterogeneity analysis. Radiomic features were extracted using PyRadiomics. Feature selection was performed using t-test filtering, Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression within the training folds. Multiple machine learning classifiers were constructed using three-fold cross-validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, calibration curves, and decision curve analysis (DCA).ResultsOf the 160 lesions, 51 (31.9%) were malignant and 109 (68.1%) were benign. The habitat radiomics model outperformed the clinical-ultrasound model. The optimal multilayer perceptron classifier achieved an AUC of 0.910 and an accuracy of 0.869 in the validation cohort. Decision curve analysis demonstrated higher net benefit across a range of threshold probabilities, and calibration curves indicated good agreement between predicted and observed outcomes.ConclusionMultimodal habitat radiomics integrating ABVS and conventional ultrasound demonstrated promising performance for risk stratification of biopsy-selected BI-RADS 4A lesions and may provide supportive information for individualized clinical decision-making. Further prospective multicenter validation is warranted before clinical application.

BackgroundMRI-PDFF enables accurate non-invasive quantification of liver fat, but its routine use is limited by cost, availability, and workflow constraints. Ultrasound-derived fat fraction (UDFF) may offer a more accessible quantitative alternative, yet its added value over the conventional hepatorenal index (HRI) remains uncertain when both are judged against the same MRI-PDFF reference standard.MethodsThis secondary cross-sectional analysis included 137 participants from a prospectively collected three-center cohort with MRI-PDFF, UDFF, and evaluable HRI measurements. Hepatic steatosis was defined as MRI-PDFF ≥5.0%. Diagnostic performance was compared using ROC analysis and paired DeLong testing, with center-stratified and alternative MRI-PDFF threshold analyses. Measurement reproducibility was evaluated for both UDFF and HRI.ResultsMRI-PDFF-defined hepatic steatosis was present in 85 of 137 participants. UDFF showed higher diagnostic discrimination than HRI (AUC, 0.945 [95% CI, 0.907–0.983] vs. 0.779 [95% CI, 0.700–0.859]; paired DeLong p = 0.0002), with consistently high center-specific AUCs (0.910–0.960) and similar results across MRI-PDFF thresholds of 5.2, 5.5, and 6.0%. UDFF demonstrated excellent within-examination repeatability and good interobserver reproducibility (ICC, 0.983 and 0.884, respectively); HRI also showed good intraobserver and interobserver reproducibility (ICC, 0.889 and 0.844, respectively).ConclusionIn this three-center cohort, UDFF provided stronger and more consistent MRI-PDFF-referenced discrimination for hepatic steatosis than HRI, while both methods showed good measurement reproducibility. UDFF may therefore provide a more robust quantitative ultrasound approach for hepatic steatosis assessment.

This study aims to evaluate the effectiveness of combining transvaginal ultrasound (US)-based radiomics and deep learning model for the accurate differentiation between benign and malignant ovarian tumors in large-scale studies. A multicenter retrospective study collected grayscale and color US images of ovarian tumors. Patients were divided into training, internal, and external validation groups. Models including a convolutional neural networks (CNN), optimal radiomics, and a combined model were constructed and evaluated for predictive performance using area under curve (AUC), sensitivity, and specificity. The DeLong test compared model AUCs with O-RADS and expert assessments. 3193 images from 2078 patients were analyzed. The CNN achieved AUCs of 0.970 (internal) and 0.959 (external), respectively. Optimal radiomic model achieved AUCs of 0.949 (internal) and 0.954 (external), respectively. The combined CNN-radiomics model attained the highest AUC of 0.977 (internal) and 0.972 (external), respectively, outperforming individual models, O-RADS, and expert methods (p < 0.05). The combined CNN-radiomics model using transvaginal US images provides more accurate and reliable ovarian tumor diagnosis, enhancing malignancy prediction and offering clinicians a more precise diagnostic tool.

BackgroundConstruction and validation of an automated breast volume ultrasound (ABVS)-based nomogram for assessing axillary lymph node (ALNs) metastasis in axillary ultrasound (AUS)-negative early breast cancer.MethodsA retrospective study of 174 patients with AUS-negative early-stage breast cancer was divided into a training and test with a ratio of 7:3. Radiomics features were extracted by combining images of intra-tumor and peri-tumor ABVS. Select the best classifier from 3 machine learning techniques to build Model 1and radiomics-score (RS). Differences in ER, PR, Her-2, Ki-67 expression were analyzed for intra-tumoral and peri-tumoral habitat radiomics features. Model 2 (based on sonogram features) and Model 3 (based on RS and sonogram features) were constructed by multivariate logistic regression. Efficiency of the models was evaluated by the area under the curve (AUC). Plotting the nomogram and evaluating its treatment in ALN≥3 according to Model 2 and Model 3.ResultIntratumoral and peritumoral 5 mm radiomics features were screened using least absolute shrinkage and selection operator (LASSO), and logistic regression was used as a classifier to build the best-performing Model 1. Using unsupervised cluster analysis, intratumoral and peritumoral 5mm were classified into 3 habitats, and they differed in PR and Her-2 expression. Model 2 (combining diameter and microcalcification) and Model 3 (combining RS and microcalcification) were created by multivariate logistic regression. Model 3 achieves the highest AUC in both the training (0.827) and validation (0.768) sets. The Nomo-score was calculated based on nomogram-model2 and nomogram-model3, revealing a positive correlation between ALN burden and Nomo-score. Combined with the optimal thresholds, nomogram-model2 screened 54.6%-100% of patients with ALN ≥3 and nomogram-model3 screened 81.8%-100% of patients with ALN ≥3.ConclusionThe ABVS-based nomogram is an effective tool for assessing ALN metastasis, and it can provide a preoperative basis for individualized treatment of breast cancer.

Background: Accurate early diagnosis of ovarian cancer is crucial. The objective of this research is to create a comprehensive model that merges clinical variables, O-RADS, and deep learning radiomics to support preoperative diagnosis and assess its efficacy for sonographers. Materials and methods: Data from two centers were used: Center 1 for training and internal validation, and Center 2 for external validation. DL and radiomics features were extracted from transvaginal ultrasound images to create a DL radiomics model using the LASSO method. A machine learning model ensemble was created by merging clinical variables, O-RADS scores, and DL radiomics model predictions. The model's effectiveness was evaluated by measuring the area under the receiver operating characteristic curve (AUC) and analyzing its impact on improving the diagnostic skills of sonographers. Moreover, the model's additional usefulness was assessed through integrated discrimination improvement (IDI), net reclassification improvement (NRI), and subgroup analysis. Results: The ensemble model demonstrated superior diagnostic performance for ovarian cancer compared to standalone clinical models and clinical O-RADS models. Notably, there were significant improvements in the NRI and IDI across all three datasets, with p-values < 0.05. The ensemble model exhibited exceptional diagnostic performance, achieving AUCs of 0.97 in both the internal and external validation sets. Moreover, the implementation of this ensemble model substantially improved the diagnostic precision and reliability of sonographers. The sonographers' average AUC improved by 11 % in the internal validation set and by 7.7 % in the external validation set. Conclusions: The ensemble model significantly enhances preoperative ovarian cancer diagnosis accuracy and improves sonographers' diagnostic capabilities and consistency.

Qin, Xiachuan; Xiao, Weihan; Zhou, Wang; Wang, Junli; Ye, Xianjun; Ren, Tiantian; Zong, Liang; Xiu, Xiaoling; Long, Qiongxian; Yuan, Hongmei; Zhao, Junjie; Wen, Yanting; Guo, Xiaoguang; He, Fanding; Zhang, Chaoxue Author Information

Those who have severe fibrosis (F2 ≥ 2 stage) are at the greatest risk for the advancement of the illness among non-alcoholic fatty liver patients. To forecast the non-alcoholic steatohepatitis (NASH) probability accompanied by significant fibrosis, we propose to develop and validate a nomogram liver imaging reporting and data system, providing robust evidence for preventing and treating clinical liver diseases. The study used SD rats to create a model of hepatic steatosis and fibrosis by feeding them a high-fat diet and injecting Ccl4 subcutaneously. Radiomics characteristics were derived from two-dimensional liver ultrasound images of the rats, and a radiomics model was constructed, with rad-scores calculated accordingly. Univariate and multivariate logistic regression was employed to ascertain the clinical characteristics of rats and liver elasticity values, aiming to establish a clinical model. Ultimately, a clinical radiomics model was created by integrating the rad-score from the radiomics model with independent clinical characteristics from the clinical model. A forest plot was generated to depict this integration. The forest plot's performance was assessed by the use of the area under the receiver operating characteristic (ROC) curve (AUC), decision curve analysis, and calibration curve. The areas under the receiver operating characteristic curve (AUC) for the training set and validation set of the clinical radiomics model were 0.986 and 0.971, respectively. Decision curve analysis showed that the clinical radiomics model had the highest net benefit across most threshold probability ranges. The nomogram and clinical radiomics model, which consists of clinical characteristics, real-time shear wave elastography, and radiomics, provide excellent predictive capability in assessing the likelihood of fibrotic NASH.

Predicting low nuclear grade DCIS before surgery can improve treatment choices and patient care, thereby reducing unnecessary treatment. Due to the high heterogeneity of DCIS and the limitations of biopsies in fully characterizing tumors, current diagnostic methods relying on invasive biopsies face challenges. Here, we developed an ensemble machine learning model to assist in the preoperative diagnosis of low nuclear grade DCIS. We integrated preoperative clinical data, ultrasound images, mammography images, and Radiomic scores from 241 DCIS cases. The ensemble model, based on Elastic Net, Generalized Linear Models with Boosting (glmboost), and Ranger, improved the ability to predict low nuclear grade DCIS preoperatively, achieving an AUC of 0.92 on the validation set, outperforming the model using clinical data alone. The comprehensive model also demonstrated notable enhancements in integrated discrimination improvement and net reclassification improvement (p < 0.001). Furthermore, the Radiomic ensemble model effectively stratified DCIS patients by risk based on disease-free survival. Our findings emphasize the importance of integrating Radiomic into DCIS prediction models, offering fresh perspectives for personalized treatment and clinical management of DCIS.

INTRODUCTION:Effective prediction methods for monitoring the risk of recurrence of colorectal cancer liver metastasis (CRLM) and guiding postoperative adjuvant treatment (ACT) are currently lacking. This study aimed to evaluate the value of postoperative dynamic circulating tumor DNA (ctDNA) monitoring in guiding ACT and predicting recurrence compared with other clinicopathological factors. METHODS:This prospective study enrolled 414 consecutive patients with CRLM who underwent radical resection. Regular dynamic ctDNA monitoring was performed every 3 months until 1 year postoperatively or on clinical recurrence. ctDNA detection was performed using the J25 detection panel, previously constructed and verified at our center. RESULTS:Postoperative ctDNA status at 1 month [hazard ratio (HR) = 3.64, 95% confidence interval (CI) 2.37-5.59, P <0.0001] significantly distinguished long-term survival. ctDNA-positive patients (HR = 0.228, 95% CI 0.116-0.446, P <0.0001) benefitted more from ACT than ctDNA-negative patients ( P = 0.39). Dynamic assessment of ctDNA status at 1 and 3-6 months postoperatively showed a significantly better prognosis in patients who remained negative or converted to negative than in those who remained positive (HR = 11.20, 95% CI 3.90-32.22, P <0.0001) or negative turned positive (HR = 3.61, 95% CI 1.31-9.98, P = 0.023). ctDNA status post-ACT and 1-month post-surgery exhibited higher area under the receiver operating characteristic (AUROC = 0.73 and 0.71) for recurrence-free survival than that of postoperative carcinoembryonic antigen (AUROC = 0.53) or pathological tumor regression grade (AUROC = 0.55). CONCLUSION:The J25 panel showed good performance for dynamic monitoring of ctDNA levels after radical resection of CRLM, improving static prognosis prediction, dynamic recurrence monitoring, and further guidance on ACT and early recurrence intervention.

The aim of this study was to develop a combined deep-learning model utilizing liver ultrasound, liver elastography images, and clinical features to predict and diagnose fibrotic non-alcoholic steatohepatitis (NASH). A rat model of liver steatosis and fibrosis was established through a high-fat diet and subcutaneous CCl₄ injections. Two-dimensional ultrasound and shear wave elastography (SWE) images were acquired. Three deep learning models, based on the ResNet-18 architecture, were designed: (1) a pure image model using only liver ultrasound, (2) a pure image model using only liver elastography, and (3) a combined model incorporating liver ultrasound, liver elastography images, and clinical features. The performance of these models was evaluated using three-fold cross-validation, receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and calibration curves. The combined deep learning model demonstrated the highest area under the curve (AUC) of 0.879. DCA revealed that the multimodal model provided superior net benefits across most threshold probability ranges for predicting and diagnosing fibrotic NASH. The combined deep learning model based on the ResNet-18 architecture exhibits promising performance in predicting and diagnosing fibrotic NASH.

Dear Editor, Intrahepatic cholangiocarcinoma (ICC) is a malignant tumour originating from the epithelial cells of the intrahepatic bile ducts. In recent years, its incidence has shown an upward trend globally. Notably, hepatitis B virus (HBV) infection is one of the significant risk factors for ICC.1 Despite significant advancements in medical imaging and molecular biology technologies, predicting the prognosis of HBV-associated ICC patients remains challenging. One major reason for this challenge is the complex interactions between HBV infection, genetic mutations and tumour behaviour, which increase the uncertainty of prognosis predictions. As a result, traditional single indicators are insufficient for comprehensively assessing patient outcomes. Radiomics is a technology that extracts a large number of quantitative features from medical images, capturing the spatial structure and morphological changes of tumours.2 Genomics, on the other hand, focuses on deciphering DNA sequence information, revealing the contributions of genetic variations to disease development. This study aims to develop and validate a predictive model that integrates radiomic features with genomic information. By doing so, it seeks to overcome the limitations of existing biomarkers, better meet the needs for personalised treatment of HBV-associated ICC patients and provide valuable references for future research and clinical practice. A total of 389 intrahepatic cholangiocarcinoma (ICC) patients were retrospectively included and divided into a training cohort (210 patients), an internal validation cohort (90 patients) and an external validation cohort (89 patients). Table S1 displayed the clinical and imaging characteristics. The results showed that most clinical characteristics did not differ significantly between the groups, including age (p = .188), gender distribution (p = .456), the proportion of ferritin ≤323 (p = .282), the proportion of high PIVKA-II (p = .988), HBV infection rate (p = .158), perineural invasion (p = .294) and AJCC 8th edition Classification of Malignant Tumors (TNM) staging (p = .455). The only characteristic that showed a statistically significant difference was the presence of vessel cancer embolus (VCE), with the training cohort (19.8%) significantly higher than the internal validation cohort (7.4%) and the external validation cohort (10.4%), p = .044. There were no extreme biases between the three cohorts in baseline data. Figure 1 showed the flow diagram of the exclusion criteria of the ICC radiomic datasets. A total of 972 features of magnetic resonance imaging (MRI) images were extracted from the ROIs using the PyRadiomics Python package,3 and those with ICC values > 0.8 on both intra- and inter-observer agreement analyses were retained. Table S2 comprehensively summarised the essential patient demographics, such as age, gender and tumour size, along with critical quantitative characteristics extracted from imaging data, including tumour volume and textural features. In our study, we evaluated multiple machine learning models to identify the best-performing algorithm for predicting ICC patient outcomes. Based on our analysis, random forest (RF) emerged as the top performer due to its ability to handle high-dimensional data and capture complex interactions, which are common in genomics and proteomics datasets. Its robustness against overfitting and provision of feature importance scores were particularly advantageous for our multi-omics approach. While support vector machines (SVM) showed strong performance, especially in high-dimensional spaces, the computational cost and interpretability challenges made it less favourable for our specific application. Logistic regression (LR), although simpler and more interpretable, did not capture the complexity of the data as effectively as RF. In the training cohort, radiomic models were capable of predicting the pathological factors, including HBV infection, cell differentiation, CA19-9 and neuro invasion (Figures 2A and S1A). Overall, the RF model performed the best in the training cohort, but its performance in the external validation cohort may have been affected by changes in data distribution. To address this issue, we have added a new external validation cohort 2 containing 202 patients. This larger sample size helped to better capture the underlying data distribution and reduce the risk of overfitting. Additionally, we have refined our analysis methods to include additional regularisation techniques and hyperparameter tuning, enhancing the robustness and reliability of our model. As a result, the area under the curve (AUC)values of the RF model for the new external validation cohort have shown some improvement, although they were still lower than those in the training cohort. This suggested that while overfitting was a concern, the model's performance can be improved with a larger and more diverse validation set. Detailed information including threshold, sensitivity, specificity, accuracy, precision and evaluation statistics from three model construction were presented in Table S3. Subsequently, we delved into whether radiomic models could predict somatic mutations in the most frequently mutated genes in ICC, specifically KRAS, BRAF, FGFR2 and IDH1/2.4 Our results highlighted that among these, FGFR2 mutations were predicted with the utmost accuracy, achieving an AUC value of 0.86 (Figure S1B). While the predictive performance for other genes, such as KRAS and BRAF, may not have been as prominent as that for IDH1/2 and FGFR2, which both achieved an AUC value of 0.86. Despite its moderate performance in predicting specific gene mutations such as KRAS and FGFR2, the radiomics-genomics model retains significant clinical utility. When the model indicates a higher likelihood of these mutations, clinicians can incorporate this information into their treatment planning processes and refine individualised therapy strategies following comprehensive genetic testing. Moreover, radiomics features associated with these gene mutations can serve as auxiliary tools for evaluating ICC patient prognosis. We conducted a comprehensive genetic and proteomic characterisation of ICC samples based on HBV infection status. The results highlighted the prevalence of genetic variations among 55 ICC samples, with 96.36% exhibiting variations, including missense mutations, in-frame deletions and insertions (Figure 2B). The most frequently detected driver gene variants associated with HBV status were USP17L7, MUC4, USP17L2, ZNF99, HNRNPL2, NBPF12, HCAR2 and TRIM49 (Figure S1C). The Table S4 presented a comparison of variant frequencies between HBV-negative and HBV-positive samples. While there were notable differences in variant frequencies for some genes, such as USP17L7 and HCAR2, between the two groups, the majority of genes exhibit statistically insignificant differences in their variant frequencies. In the TCGA database, the mutation rate of the MUC4 gene was recorded at 21%. MUC4 was a mucin whose aberrant expression was associated with various cancer types, including ICC.5 The high mutation rate suggested that MUC4 may play a significant role in the pathogenesis of ICC (Figure S1D). In addition, ICC patients harbouring mutations in these genes exhibited higher tumour mutation burdens and mutation counts,6, 7 which suggested that mutations in these genes may be closely linked to genetic instability and the process of tumour evolution in ICC (Figure S1E). Proteomic-based gene set enrichment analysis revealed that oncogenic pathways, including DNA replication, cell–cell junction organisation, cell cycle and DNA repair, were significantly upregulated in HBV-positive ICC (Figures 2C and S1F). Particularly, the expression of DNA repair-related genes PARP1 and TOP1 was identified exclusively in HBV-positive tumours, which were validated through immunohistochemistry (Figure 2D). The upregulated DNA repair pathways may render HBV-positive ICC cells dependent on specific repair mechanisms. For example, PARP inhibitors have been successfully used in BRCA1/2-deficient tumours, and this strategy can be extended to other cancer types with upregulated DNA repair pathways. The effects of novel damage response (DDR) inhibitors, such as ataxia telangiectasia mutated (ATM) kinase inhibitor, on HBV-positive ICC could be tested using preclinical models or clinical trials. Differential gene expression analysis identified 672 proteins that were significantly over-represented in HBV-positive ICCs (fold change > 2; adjusted p < .05; Table S5). To gain insights into the potential relationship between HBV status and clinical pathological as well as radiomic features, we applied the correlation analysis. As presented in Table S6, some radiomic features, including shape elongation and certain parameters of the grey-level co-occurrence matrix, exhibited significant correlations with HBV status, providing data supported for further exploration of the impact of HBV on imaging features. To further interpret the radiomics models pertaining to HBV, we presented two case examples of ICC patients. Our model accurately predicted the HBV status of these patients (Figure 2E). The predictions were based on a set of radiomic features extracted from the segmented MRI images. Specifically, these features included characteristics such as shape elongation and certain parameters of the grey-level co-occurrence matrix, which have been shown to differ significantly between HBV-infected and non-infected individuals. By leveraging these distinct imaging biomarkers, the model can differentiate between HBV-positive and HBV-negative statuses. Next, we applied the RF classifier to predict HBV status on the basis of tumour radiomic features. Our radiomic models demonstrated a significant ability to distinguish HBV-positive from HBV-negative status (training cohort AUC = 0.73). However, to provide a more comprehensive assessment of the model's performance, we also evaluated it using an independent internal validation cohort (n = 90; AUC = 0.64) and an external validation cohort (n = 89; AUC = 0.65). These results indicated that, although there was a minor reduction in performance, the model retained its predictive capability across different patient populations (Figure S1G). We also provided a comprehensive baseline statistical analysis of relevant imaging features in samples stratified by HBV infection status (Table S7). In the training cohort, significant differences were observed in multiple imaging features between HBV-infected and non-infected individuals, including Elongation, Maximum 2D Diameter Column, GLCM Correlation and GLCM Idm. However, the p values for these features in the internal and external validation cohorts, including a new external validation set of 202 cases, did not show significant improvement, suggesting that further validation was needed to confirm their robustness and generalisability. This predictive outcome was consistent with clinical test results, indicating that radiomics features can effectively reflect the HBV infection status of tumours, offering the potential for non-invasive diagnosis of HBV-related tumours. HBV infection may affect the metabolic pathways of hepatocytes, leading to abnormalities in glycogen synthesis and breakdown, as well as fat metabolism.8 These changes may manifest radiologically as hepatic steatosis or inhomogeneity in density.9 Here, we found that radiomic features were associated with cancer-related KEGG pathways covering multiple aspects of the cancer molecular system (Figure 2F). Specifically, transcriptional activity of several molecular signalling pathways, including chemical carcinogenesis, drug metabolism was negatively associated with tumour size features, indicating that they were more active in smaller tumours than larger tumours. In addition, MUC4 and USP17L7 mutations were highly correlated with HBV status, and patients with HBV-infected ICC often exhibited low LRBA, MUC5AC and MUC1 expression (Figure S1H). In order to evaluate the predictive capability of a machine learning-based radiomics model, we extracted MRI data from a cohort of ICC patients within an external validation set and identified cases predicted to harbour gene mutations, yielding a total of 16 subjects (Figure 2G). Subsequent WES analysis revealed that the frequencies of gene mutations among these 16 patients in the validation set were as follows: MUC4: 94%, NBPF12: 94%, USP17L7: 88% and TRIM49: 88%. These results demonstrated the promising potential of this radiomics model in accurately predicting gene mutations without invasive procedures, warranting further investigation on a larger scale to validate its efficacy. We next conducted scRNA-seq experiments on samples from three ICC patients, comprising one HBV-ICC patient (ICC 1) and two non-HBV-ICC (ICC 2 and ICC 3) patients. By examining cell-type-specific markers, tumour-associated markers and copy number variations (CNVs) within individual cells, we were able to definitively identify the specific types of different cells (Figure 3A). Compared with HBV-negative samples, HBV-positive samples have fewer cell–cell interaction types and stronger interaction strength (Figures 3B and S2A) Detailed changes in the number and strength of intercellular interactions among samples with different HBV statuses were shown in Figure 3C,D. Nine functional groups were significantly enriched in CD4+ and CD8+ T cells (p ≤ .05), including somatic cell DNA recombination, regulation of DNA recombination, DNA recombination and DNA deamination (Figure 3C), which was consistent with the results of GO term pathway analysis (Figure 3E). The observed low and significant expression of GSTP1 and TOP1, along with a similar but not significant trend for ATM, suggested alterations in various processes critical for maintaining genomic stability (Figures 3F and S2B). Next, the pseudotime analysis revealed consistently low expression levels of GSTP1, TOP1 and ATM, with only slight increase observed over time in HBV-positive ICC (Figure 3G). These findings provided valuable insights into the molecular mechanisms and cellular interactions that distinguish HBV-ICC from non-HBV-ICC, highlighting the importance of CD8+ T cells in specific molecular processes and the potential roles of GSTP1, TOP1 and ATM in DNA replication and repair within HBV-ICC. Three types of OS models were developed, including a pathology-based model, an imaging-pathology based model and an imaging-pathology-genomic based model (Figure S3A,B). The imaging-pathology based OS model also included several predictors, including Radscore derived from arterial phase imaging, pT stage, pN stage, CA125, CA199 and CEA level, with a model cutoff value of 193.49. Based on the cutoff value of 193.49, this model categorised ICC patients into a high-risk group (median value of Risk2: 304.1, interquartile range (IQR): 210–447) and a low-risk group (median value of Risk2: 142.4, IQR: 67–165.89) (p < .0001) (Figure 4A). Finally, the calibration curve of the imaging-pathology-based OS model demonstrated good agreement between predictions and observations (Figure 4B). In predicting the OS of patients with ICC, the imaging-pathology based OS model demonstrated the best performance. Specifically, this model had the highest C-index and the lowest p-value, showcasing its superior capability in predicting patient prognosis. Utilising radiomics derived from MRI, whole-transcriptome data and machine learning techniques, this is the first study combining radiomic features from MRI images with full-genome measurements that depict the multi-layered tumour molecular systems in ICC. Our radiogenomics nomogram provides a powerful and non-invasive method for predicting ICC prognosis, thereby supporting more informed clinical decision-making and facilitating personalised treatment approaches. The future integration of radiomics with deep learning holds significant potential to revolutionise medical diagnostics and personalised medicine. By leveraging multi-parametric imaging frameworks, such as MPRAD, this integration combines information from multiple imaging sequences to provide more comprehensive diagnostic data. Deep learning algorithms can automatically uncover complex patterns within these images, thereby enhancing predictive accuracy and aiding in the discovery of subtle disease indicators. Future research can extend to larger and more diverse ICC patient populations to provide more robust model validation, and through multi-centre collaborations, achieve the integration of resources and expertise, ensuring the reproducibility of results. Despite the multiple measures we took in the study design and data analysis to ensure the reliability and validity of the results, several limitations remain in this study. First, there is a discrepancy in the incidence of VCE between the training and validation sets, which may affect the model's generalisability. This difference is likely due to the inherent variability in patient recruitment periods and sources. Future studies will aim to reduce these differences by increasing the sample size and using more consistent recruitment criteria. Besides, the moderate AUC of HBV prediction observed in external validation indicates that while our model performs reasonably well within its original context, its generalisation across different populations or settings may be less effective. This limitation can be attributed to data heterogeneity arising from differences in data collection methods and patient demographics between the training and external validation sets, or cohort variability due to variations in clinical characteristics and treatment regimens. To enhance the model's generalisability in future research, we plan to incorporate more diverse training data from multiple centres and regions to capture a broader spectrum of variability, thus enhancing the model's robustness. Y. Jia, M. Wan, Y. Shen, and J. Wang analyzed the data in the study and, as co-first authors, contributed equally to this work. X. Luo and M. He participated in the data collection and provided critical feedback that shaped the final research outcomes. R. Bai and W. Xiao contributed to the data collection and offered essential feedback. X. Zhang and J. Ruan designed and supervised the study and provided critical feedback at all stages of the research. All authors provided critical feedback and helped shape the research, analysis and manuscript. This work was supported by the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (2024C03175), Beijing Science and Technology Innovation Medical Development Foundation (KC2023-JX-0186-FZ099), National Natural Science Foundation of China (82473004, 81874173), Zhejiang Provincial Natural Science Foundation of China (LY22H160019, MS25H160091), and Beijing Xisike Clinical Oncology Research Foundation (Y-MSDZD2022-0161). The authors declare no conflict of interest. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

BACKGROUND:Accurate preoperative assessment of axillary lymph node metastasis (ALNM) in breast cancer is crucial for guiding treatment decisions. This study aimed to develop a deep-learning radiomics model for assessing ALNM and to evaluate its impact on radiologists' diagnostic accuracy. METHODS:This multicenter study included 866 breast cancer patients from 6 hospitals. The data were categorized into training, internal test, external test, and prospective test sets. Deep learning and handcrafted radiomics features were extracted from ultrasound images of primary tumors and lymph nodes. The tumor score and LN score were calculated following feature selection, and a clinical-radiomics model was constructed based on these scores along with clinical-ultrasonic risk factors. The model's performance was validated across the 3 test sets. Additionally, the diagnostic performance of radiologists, with and without model assistance, was evaluated. RESULTS:The clinical-radiomics model demonstrated robust discrimination with AUCs of 0.94, 0.92, 0.91, and 0.95 in the training, internal test, external test, and prospective test sets, respectively. It surpassed the clinical model and single score in all sets (P < .05). Decision curve analysis and clinical impact curves validated the clinical utility of the clinical-radiomics model. Moreover, the model significantly improved radiologists' diagnostic accuracy, with AUCs increasing from 0.71 to 0.82 for the junior radiologist and from 0.75 to 0.85 for the senior radiologist. CONCLUSIONS:The clinical-radiomics model effectively predicts ALNM in breast cancer patients using noninvasive ultrasound features. Additionally, it enhances radiologists' diagnostic accuracy, potentially optimizing resource allocation in breast cancer management.

Background: It is difficult to make a definite diagnosis of borderline epithelial ovarian tumors before surgery. In order to avoid incorrectly classifying tumors as benign, a differential diagnosis model was developed to distinguish between benign and borderline epithelial tumors utilizing multimodal information. Method: A multicenter study was conducted. A retrospective analysis of the transvaginal ultrasonography and clinical data of patients who underwent surgery and received pathological diagnoses of borderline and benign epithelial ovarian tumors was conducted. Both Univariate and multivariate logistic regression analyses were used to develop a diagnostic model for borderline epithelial tumors. The efficacy and feasibility of this model were assessed through examination of training, internal validation, and external test sets. Results: There was a significant difference in D-dimer levels between borderline and benign epithelial tumors. Abnormal CA125, D-dimer, maximum mass diameter > 10 cm, regular and irregular solid portions, and blood flow in the mass were independent risk factors for borderline epithelial ovarian tumors. The diagnostic model was evaluated by the Hosmer–Lemeshow test and demonstrated strong fitting capabilities. ROC curve analysis of the training set, verification set, and external test set confirmed the model’s predictive ability. Conclusions: These independent risk factors may be combined to assess the risk of borderline epithelial ovarian tumors. Our findings will assist novice gynecologic sonographers in distinguishing between benign and borderline epithelial tumors.

Background:The automated breast volume scanner (ABVS), a type of ultrasound device, plays a crucial role in breast cancer screening; however, the ABVS data volume places a strain on clinicians. We aimed to develop an artificial intelligence (AI) model for the detection and classification of lesions as benign or malignant during ABVS examination. Methods:This retrospective study included 1,284 patients with 1,769 lesions who underwent ABVS examination between January 2017 and August 2021. The lesions were randomly divided into training and test sets at a 7:3 ratio. Using the test set, the performance of the You Only Look Once (YOLO) AI model, based on the YOLO version 8 architecture, in single-target (background vs. lesion), categorical (benign vs. malignant), and varied lesion diameter detection was evaluated. Finally, differences in the diagnoses of four radiologists with different levels of experience before and after receiving AI model assistance were assessed. Results:The recall of the YOLO AI model for single-target detection was 0.983. The precision, recall, mean average precision (mAP) 50, and F1-score of the YOLO AI model for categorized target detection were 0.887, 0.866, 0.919, and 0.876, respectively. While the precision, recall, mAP50, and F1-score of the YOLO AI model for the classification of lesions with diameters ≤10 mm, 10 mm < diameters ≤ 20 mm, 20 mm < diameters ≤ 30 mm, and diameters >30 mm were 0.910, 0.806, 0.868, 0.855; 0.895, 0.844, 0.911, 0.869; 0.876, 0.867, 0.917, 0.871; and 0.882, 0.898, 0.941, 0.890, respectively. The area under the curve (AUC) values of the radiologists after they received YOLO AI assistance in the diagnosis of breast lesions were 0.806, 0.890, 0.897, and 0.895, respectively, and these AUC values were better than their AUC values before they received YOLO AI assistance (P<0.001). Conclusions:The YOLO AI model can effectively identify and characterize breast lesions. It improves radiologists' diagnostic performance and bridges expertise gaps between radiologists.

RATIONALE AND OBJECTIVES:The aim of this study was to develop a deep learning radiomics nomogram (DLRN) based on B-mode ultrasound (BMUS) and color doppler flow imaging (CDFI) images for preoperative assessment of lymphovascular invasion (LVI) status in invasive breast cancer (IBC). MATERIALS AND METHODS:In this multicenter, retrospective study, 832 pathologically confirmed IBC patients were recruited from eight hospitals. The samples were divided into training, internal test, and external test sets. Deep learning and handcrafted radiomics features reflecting tumor phenotypes on BMUS and CDFI images were extracted. The BMUS score and CDFI score were calculated after radiomics feature selection. Subsequently, a DLRN was developed based on the scores and independent clinic-ultrasonic risk variables. The performance of the DLRN was evaluated for calibration, discrimination, and clinical usefulness. RESULTS:The DLRN predicted the LVI with accuracy, achieving an area under the receiver operating characteristic curve of 0.93 (95% CI 0.90-0.95), 0.91 (95% CI 0.87-0.95), and 0.91 (95% CI 0.86-0.94) in the training, internal test, and external test sets, respectively, with good calibration. The DLRN demonstrated superior performance compared to the clinical model and single scores across all three sets (p < 0.05). Decision curve analysis and clinical impact curve confirmed the clinical utility of the model. Furthermore, significant enhancements in net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indicated that the two scores could serve as highly valuable biomarkers for assessing LVI. CONCLUSION:The DLRN exhibited strong predictive value for LVI in IBC, providing valuable information for individualized treatment decisions.

Background A reliable assessment of hepatic steatosis is imperative for the effective management of metabolic dysfunction-associated steatotic liver disease (MASLD). This study assesses the effectiveness of ultrasound-derived fat fraction (UDFF) in measuring hepatic steatosis and determines diagnostic thresholds for different severity levels. Methods This prospective cross-sectional study involved 79 participants (mean age 42.8 ± 13.8 years) recruited from two centers. MRI proton density fat fraction (PDFF) served as the reference standard for assessing hepatic steatosis. Pearson correlation coefficients were applied to determine the relationship between UDFF and MRI-PDFF, while Bland-Altman analysis evaluated the measurement consistency between UDFF and PDFF.ROC curve analysis evaluated the diagnostic performance of UDFF against visual score, Fatty Liver Index (FLI), the CAP score (CAPS), and Hepatic Steatosis Index (HSI). Results The UDFF showed a strong correlation with MRI-PDFF (r = 0.84, p < 0.001), with a mean bias of 2.06% and 95% limits of agreement ranging from − 7.03–11.15%. The AUC values for UDFF in diagnosing steatosis grades ≥ S1, ≥S2, and S3 were 0.95, 0.97, and 0.94, respectively, outperforming visual score, FLI, CAPS, and HSI. The optimal UDFF cutoff values for these grades were 8.5%, 16.5%, and 22%. Conclusion UDFF shows high consistency in diagnostic performance with PDFF and steatosis grades, although UDFF values tend to be slightly higher than those of PDFF.

Background Non-Alcoholic Steatohepatitis(NASH) is a crucial stage in the progression of Non-Alcoholic Fatty Liver Disease(NAFLD). The purpose of this study is to explore the clinical value of ultrasound features and radiological analysis in predicting the diagnosis of Non-Alcoholic Steatohepatitis. Method An SD rat model of hepatic steatosis was established through a high-fat diet and subcutaneous injection of CCl4. Liver ultrasound images and elastography were acquired, along with serum data and histopathological results of rat livers.The Pyradiomics software was used to extract radiomic features from 2D ultrasound images of rat livers. The rats were then randomly divided into a training set and a validation set, and feature selection was performed through dimensionality reduction. Various machine learning (ML) algorithms were employed to build clinical diagnostic models, radiomic models, and combined diagnostic models. The efficiency of each diagnostic model for diagnosing NASH was evaluated using Receiver Operating Characteristic (ROC) curves, Clinical Decision Curve Analysis (DCA), and calibration curves. Results In the machine learning radiomic model for predicting the diagnosis of NASH, the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve for the clinical radiomic model in the training set and validation set were 0.989 and 0.885, respectively. The Decision Curve Analysis revealed that the clinical radiomic model had the highest net benefit within the probability threshold range of > 65%. The calibration curve in the validation set demonstrated that the clinical combined radiomic model is the optimal method for diagnosing Non-Alcoholic Steatohepatitis. Conclusion The combined diagnostic model constructed using machine learning algorithms based on ultrasound image radiomics has a high clinical predictive performance in diagnosing Non-Alcoholic Steatohepatitis.